Anthropic to embed invisible watermarks in text and files generated by Claude
The Facts
- Anthropic said Claude will embed a machine-readable watermark into text its models generate.
- Anthropic will also attach digitally signed provenance metadata to supported generated files.
- Anthropic linked the change to the EU AI Act's Article 50(2) transparency code of practice, which it signed.
- Claude models launched on or after Aug. 2, 2026 support the marking from day one.
- Anthropic is applying the marking globally, not only in the European Union.
- Anthropic says the text mark is designed to survive copying and minor edits.
- Anthropic has not given a timeline for adding the marking to older Claude models.
- Users who say they use Claude to proofread their own writing objected to the change online.
Context
How does an invisible watermark work inside ordinary text?
Claude subtly shifts the probabilities of the words, or tokens, it selects while writing, producing a statistical pattern that software can later look for Tom's Guide,PlayGround.ru. Anthropic says the signal does not change the meaning, quality or readability of a response NY Post,Exame. For images and other files, the company uses digitally signed provenance metadata based on the C2PA standard PlayGround.ru,Business Standard.
Does a watermark prove that AI wrote something?
No. A watermark signals only that content passed through Claude, not that the model wrote every word or that a person did not do most of the work Hindustan Times. Short passages cannot be marked reliably, according to reporting on Anthropic's documentation PlayGround.ru. Anthropic itself has listed limitations of the system CartaCapital.
What does the EU law require, and what happens if a company does not comply?
Article 50 of the EU AI Act became enforceable on Aug. 2 and requires generative AI providers to mark outputs in machine-readable formats so downstream users, regulators and platforms can identify AI-generated content Forbes,elcolombiano.com. Forbes reports that fines for non-compliance can reach €15 million or 3% of a company's total worldwide annual turnover, whichever is higher Forbes.
Where Left and Right agree, and where they split
- Where Left and Right agree
- Both read the global rollout, not the EU code, as the real event: a commitment made for one jurisdiction now sets disclosure terms for every Claude user, incompletely.
- Where Left and Right split
- Whether the story is about a public's claim to know what a machine wrote, or about users outside Europe bound by a code they never consented to.
How left and right read it
Whether ordinary people can tell what a machine wrote should not depend on which market they happen to live in, so it matters that Anthropic is applying its watermark and signed provenance metadata globally rather than only where the EU transparency code it signed applies. But the protection has a hole: only models launched on or after Aug. 2, 2026 carry the marking from day one, and there is still no timeline for older ones. Transparency owed to the public should be finished, not phased.
A rule written for Brussels ought to govern Brussels. Yet because Anthropic signed the EU AI Act's Article 50(2) transparency code, it now embeds a machine-readable mark in Claude's text and signed provenance metadata in files globally rather than only in Europe — and that mark is built to survive copying and minor edits. So users who lean on Claude to proofread their own writing object. Scope should follow consent.
“Dario Amodei's Anthropic is adding watermarks to AI-generated content in order to comply with a strict new transparency law in Europe -- drawing gripes from some users.” — New York Post
The receipts — all 27 sources
Independent coverage (27)
Facts first. Then every angle.
The day’s biggest stories in one short brief — the facts everyone agrees on, then the competing values behind the headlines. Free in your inbox.